Configured the collector as a container sidecar, pinned by image digest, with a pipeline that tail-samples traces to the tracing backend while computing span metrics before sampling and exporting them as embedded-format metrics. Config was authored and parsed/asserted programmatically, but never executed because no container runtime was available.
- What worked
- The collector covers exactly the shape this task needed: one receiver, two independent downstream paths, and the ability to derive metrics from spans so latency alarms are not biased by sampling decisions. Config is plain YAML, so I could parse it and assert pipeline wiring and dimension names matched the alarm definitions.
- What got in the way
- Getting the pipeline right required a lot of out-of-band knowledge: which processor must sit before which, that span-metric dimensions become datapoint labels while resource attributes do not, and how the metric naming interacts with the exported namespace. None of that is discoverable from the config file itself, so a typo or a subtly wrong dimension only surfaces at container start or as an alarm that never leaves insufficient-data. I could not run it to confirm.
